Meta-Learning with Shared Amortized Variational Inference

TL;DR

This paper introduces a novel variational inference method for meta-learning that maintains uncertainty over model parameters, improving over previous Monte-Carlo based approaches by using shared amortized inference networks.

Contribution

It presents a new amortized variational inference scheme for empirical Bayes meta-learning, sharing inference networks to better model uncertainty in parameters.

Findings

Outperforms previous methods on miniImageNet, CIFAR-FS, and FC100 datasets.

Prevents collapse of the conditional prior, preserving uncertainty.

Demonstrates advantages of variational approach over Monte-Carlo approximation.

Abstract

We propose a novel amortized variational inference scheme for an empirical Bayes meta-learning model, where model parameters are treated as latent variables. We learn the prior distribution over model parameters conditioned on limited training data using a variational autoencoder approach. Our framework proposes sharing the same amortized inference network between the conditional prior and variational posterior distributions over the model parameters. While the posterior leverages both the labeled support and query data, the conditional prior is based only on the labeled support data. We show that in earlier work, relying on Monte-Carlo approximation, the conditional prior collapses to a Dirac delta function. In contrast, our variational approach prevents this collapse and preserves uncertainty over the model parameters. We evaluate our approach on the miniImageNet, CIFAR-FS and FC100 datasets, and present results demonstrating its advantages over previous work.